What this game is
Plain tic-tac-toe is dead: the game is fully solved, perfect play by both sides always ends in a draw, and the whole strategy fits on a napkin. This nested version is built from the same cells, and it is not dead at all.
The board is nine small 3×3 boards arranged in a large square. The cell you take inside a small board sends your opponent to the small board in the same position: play the top-right cell and he must answer in the top-right board. If that board is already closed, he plays wherever he wants. A won small board closes and is stamped with a big mark; a board filled with no winner closes for nobody. Three small boards won in a row win the game; if all nine close without such a row, it is a draw.
The forwarding rule is what makes the game: every move is also an instruction to your opponent. Giving up a cell, even a board, to send him somewhere harmless is ordinary tactics here - it matters not only where you win but where your move sends the other player.
How the machine thinks
There is no evaluation function on this page. The machine does not know that the centre board is strong or that two marks in a row are a threat; it knows nothing about the game except the rules. Instead it runs Monte Carlo tree search: from the current position it plays out thousands of completely random games to the very end and records which first moves ended in a win more often.
The tree grows unevenly. A one-line formula (UCT) decides where the next random game starts: mostly where results have been good, sometimes where too few games have been tried. That balance between exploiting what works and exploring what is unknown is the entire mathematics of the method.
This trick - random playouts plus a search tree - is what, combined with neural networks, broke go: AlphaGo, 2016. The page How machines play tells that story. Here MCTS runs in its pure form, with no network attached: this is the only game on the site where the machine thinks this way.
About the levels
A level is thinking time and nothing else: Normal about 0.15 s per move, Hard about 0.5 s, Expert about 1.4 s. The algorithm is identical on all three; the only difference is how many random games it finishes before the clock runs out.
The small line under the board shows that number after every machine move. It is not decoration: the entire strength of the move is exactly that figure.
Is the game solved?
No. Unlike plain tic-tac-toe, the nested game has not been solved: nobody has proved how it ends under perfect play from both sides. Computer analysis and tournament practice point to an advantage for the first player - but that is an observation, not a theorem.
The first move is yours.